Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Worcester Polytechnic Institute
1
0
3
Harmful content generation in LLMs isn't a holistic property, but rather a localized phenomenon driven by specific neurons in later MLP layers acting as "harmfulness gates."